Towards hydrological extreme event monitoring using deep learning-downscaled simulated NGGM and MAGIC data and GRACE(-FO) observations

Abstract

Terrestrial Water Storage Anomalies (TWSA) derived from satellite gravimetry provide a unique, spatially integrated measure of large-scale hydrological conditions and are widely used to characterize basin-scale wet and dry states. However, their applicability to regional hydrological monitoring remains limited by their coarse spatial resolution, which hampers the detection of localized storage anomalies relevant to flood onset, evolution, and groundwater variability.Here, it is investigated an unsupervised deep-learning framework based on convolutional neural networks to enhance the spatial resolution of TWSA products across simulated and real mission scenarios. The method relies on a U-Net architecture driven by ERA5-Land hydro-climatic variables, and is first applied to end-to-end closed-loop simulations of the NGGM and MAGIC next-generation gravity missions. Particular attention is given to 5-daily solutions, which represent a key expected advancement of future gravimetric constellations compared with the monthly solutions currently provided by GRACE(-FO). In this setting, simulated TWSA fields are downscaled from 3° to 1°, while monthly 2° simulations are also considered to assess the framework under GRACE-like temporal conditions. For the syntetic products, the downscaled fields are validated using the ESA Earth System Model 3.0 as the reference signal.Results show that the downscaled products preserve consistency with native gravimetric observations while recovering finer-scale spatial structures. To assess transferability, the same pipeline is applied to real GRACE(-FO) observations, downscaling monthly 3° TWSA products into 1° maps.The framework is also extended to GRACE-derived groundwater storage anomalies (GWSA) over France, downscaled from 0.5° to 0.1° and evaluated against independent in situ groundwater observations.Finally, downscaled synthetic TWSA fields are exploited for extreme-event monitoring using threshold-based indicators derived from climatological anomalies. The analysis assesses whether the increased spatial resolution retains localized anomaly patterns otherwise smoothed at coarse scales, supporting the monitoring of flood-prone conditions.